(II.B) Basis and dimension How would you explain that a plane has two dimensions? Well, you can go in two independent directions, and no more. To make this idea precise, we formulate the DEFINITION 1. A (finite) subset { v 1,..., v m } of a vector space V (over a field F) is a basis of V if (a) they are linearly independent (over F), and (b) they span V (over F). 1 The over F means that the linear combinations are understood to be of the form α i v i, with α i F. As usual, our default will be F = R. EXAMPLE 2. A basis of R m is given by the standard basis vectors {ê i } m. Of course, one could say the same for any field: the standard basis vectors are also basis for C m, viewed as a vector space over C. But what if we view C m as a vector space over R? Then you need the 2m vectors ê 1,..., ê m and iê 1,..., iê m! (Why?) EXAMPLE 3. Let P 2 (x, y) denote the vector space of quadratic polynomials in two variables x and y, with real coefficients. This has the monomial basis {ρ 2 i (x, y)}6 = {1, x, y, xy, x2, y 2 }. Similarly, {ρ 3 i (x, y)}10 = {1, x, y, xy, x2, y 2, x 2 y, y 2 x, x 3, y 3 } gives a basis for the space P 3 (x, y) of cubic polynomials. We ll give a geometric application of this example below. 1 I am not saying that you can t have infinite bases; indeed, non-finitely-generated vector spaces (like P (x) := all polynomials) may have such a basis. I m just not defining or discussing them at this stage. 1
2 (II.B) BASIS AND DIMENSION EXAMPLE 4. Let A be an m n matrix, and V be the vector subspace of R n consisting of solutions to A x = 0. How do you compute a basis of V? By making use of the row-reduction algorithm, of course! For instance, we might have 0 0 1 1 1 1 2 0 3 0 2 4 2 4 2 0 0 1 1 0 A = rre f (A) = 2 4 3 3 3 0 0 0 0 1. 3 6 6 3 6 0 0 0 0 0 The free variables are x 2 and x 4, and an arbitrary solution is given by plugging arbitrary values in for these and solving for the pivot variables x 1, x 3, x 5. To get a basis, you plug in (x 2 = 1, x 4 = 0) and (x 2 = 0, x 4 = 1) to obtain 2 3 1 0 0, 1. 0 1 0 0 More generally, the i th basis element is given by setting the i th free variable to 1 and the rest to zero. EXAMPLE 5. Suppose A is an invertible m m matrix. I claim that its columns { c i } m are a basis for Rm. To establish this, check that: { c i } are linearly independent : γ 1 c 1 +... + γ m c m = 0 = A γ = 0 = γ = 0 (since A is invertible). { c i } span R m : show every y R m can be written y = γ 1 c 1 +... + γ m c m (= A γ). This is easy: just put γ = A 1 y. We say that V is f initely generated if some finite subset { v 1,..., v m } spans V. Here is the precise reason why dimension makes sense: PROPOSITION 6. If V is finitely generated then (i) it has a (finite) basis, and (ii) any 2 bases of V have the same number of elements.
(II.B) BASIS AND DIMENSION 3 DEFINITION 7. We define the dimension dim F (V) of V over F to be the number of elements in (ii). 2 We get (i) by going through the generating subset { v 1,..., v m } and throwing out any v i that is a linear combination of the earlier vectors. What remains is a basis. In order to show (ii) we first prove the following LEMMA 8. If v 1,..., v m span V then any independent set V has no more than m vectors. PROOF. Consider any set u 1,..., u n, n > m (of vectors in V ). Since v 1,..., v m span V, each u j is a linear combination of the v i : u j = m A ij v i, for A m n. Since m < n, there is a nontrivial solution to A x = 0, some nonzero x R n. That is, and so n x j u j = j=1 n A ij x j = 0 for each i, j=1 n m x j A ij v i = j=1 ( m n x j A ij ) v i = j=1 m 0 v i = 0. We conclude that { u j } n j=1 is not an independent set! Now let { v 1,..., v m } and { w 1,..., w n } be two bases for V. The { v i } span V and the { w j } are independent, so the Lemma = n m. The same argument with { v i } and { w j } swapped = n m, and so we ve proved the Proposition. We can rephrase the Lemma as follows: if dim(v) = m then (a) less than m vectors cannot span V, while (b) more than m vectors are linearly dependent. A Geometric Application. Given 5 points A, B, C, D, E in general position in the plane, there is a unique conic passing through 2 The F is typically omitted when there is no ambiguity. That said, in the first example we had dim C (C n ) = n while dim R (C n ) = 2n.
4 (II.B) BASIS AND DIMENSION them. Here s why: the equation of a conic Q has the form 0 = f Q (x, y) = 6 a i ρ 2 i (x, y) = a 1 + a 2 x + a 3 y + a 4 xy + a 5 x 2 + a 6 y 2, and the statement that it passes through A is 0 = f Q (x A, y A ) = a i ρ i (x A, y A ). Continuing this for B thru E we get 5 equations in 6 unknowns a i, a system which generically 3 has a line of solutions: if (a 1,..., a 6 ) solves the system then so does (c a 1,..., c a 6 ). But this just gives multiples of the same f Q, which won t change the shape of the curve, and so there is really only one nontrivial solution: through five general points in the plane there exists a unique conic Q. Similarly, given 8 points A, B, C, D, E, P 1, P 2, P 3 what can we say about the vector space of cubics passing through them (i.e. of cubic polynomials in (x,y) vanishing at all 8 points)? First of all, the vector space of all cubics 10 a i ρ 3 i (x, y) = a 1 + a 2 x + a 3 y +... + a 9 x 3 + a 10 y 3 should be thought of in terms of the coefficient vector a. The 8 constraints give a linear system 0 = 10 a iρ 3 i (x A, y A ). 0 = 10 a iρ 3 i (x 8 equations P 3, y P3 ) whose space of solutions in general (i.e., under the assumption of maximal rank[=8]) is two-dimensional there are 8 leading 1 s in rre f of the 8 10 [ρ] matrix, and therefore 2 parameters to choose freely. So start with A, B, C, D, E in general position ; we d like to construct points on the (unique) conic through them, using a straightedge. Begin by drawing the lines AB, DE, CD, and BC ; label 3 more precisely, generically means here that the matrix [ρ] (with rows indexed by A thru E, columns by i = 1 thru 6 ) is of maximal rank (= 5 ).
(II.B) BASIS AND DIMENSION 5 AB DE by P 1. (Q is drawn in a background color because we don t know what it looks like yet.) Q B C A D E BC AB CD DE P 1 Now draw (almost) any line l through A the choice is free except that l shouldn t go through B, C, D, or E. All the rest of the construction depends on this choice; in particular the final point q depends on l, and so by varying l we can vary q and in this way (we hope) construct enough points on Q to get an idea of what it looks like. Label l CD =: P 2 ; then draw P 1 P 2 and label P 1 P 2 BC =: P 3 ; finally set q := EP 3 l. We must prove q Q. Adding dotted lines to our diagram for the lines depending on l, Q B C A q E D BC AB l CD EP 3 DE P 1 P 2 P 3
6 (II.B) BASIS AND DIMENSION Now consider the following three cubic equations (where e.g. f AB (x, y) = 0 is the linear equation of the line AB, and so on): f 1 (x, y) := f Q (x, y) f P1 P 2 (x, y) = 0, f 2 (x, y) := f AB (x, y) f CD (x, y) f EP3 (x, y) = 0, f 3 (x, y) := f l (x, y) f BC (x, y) f DE (x, y) = 0. The graphical solutions of these equations are (respectively) Q P 1 P 2, AB CD EP 3, l BC DE. All 3 of these unions contain the 8 points A, B, C, D, E, P 1, P 2, P 3 ; that is, the cubic polynomials f 1, f 2, and f 3 each vanish at all 8 points. Above we explained that the (vector) space of such cubic polynomials was 10 8 = 2 dimensional; it follows that 3 vectors f 1, f 2, f 3 cannot be linearly independent and there is a nontrivial relation α f 1 (x, y) + β f 2 (x, y) + γ f 3 (x, y) = 0. Now we extract the information we want (q Q ) from this relation. First of all α cannot be zero. Otherwise β f 2 = γ f 3 are multiples of each other, which means that AB CD EP 3 = l BC DE. That would imply l was either AB, CD, or EP 3 ; and we said l cannot pass through B, C, D, or E so this gives a contradiction. So we may divide by α = f 1 (x, y) = β α f 2(x, y) γ α f 3(x, y). Now since q is contained in both EP 3 and l, f 2 (x, y) and f 3 (x, y) both vanish at q = (x q, y q ). (This is because f 2 contains a factor of f EP and f 3 contains a factor of f l.) According to the relation, therefore also f 1 = 0 at q ; this means that either f Q or f P1 P 2 vanishes at q, and so q Q or P 1 P 2. Suppose the latter: if q P 1 P 2 then the three lines P 1 P 2, EP 3, l are the same, and so E l. But once again, we disallowed such a choice of l at the outset. So q Q, and we re done. Another way of thinking of this result is that intercepts of opposite edges of a hexagon inscribed in a conic lie on line. This is a theorem proved by Pascal in 1639 known as his mystic hexagon!
Exercises EXERCISES 7 (1) Let V be a vector space over a subfield F of the complex numbers. Suppose u, v, w are linearly independent vectors in V. Prove that ( u + v), ( v + w), and ( w + u) are linearly independent. (2) Let V be the set of all 2 2 matrices A with complex entries which satisfy A 11 + A 22 = 0. (a) Show that V is a vector space over the field of real numbers, with the usual operations of matrix addition and multiplication of a matrix by a scalar. (b) Find a basis for this vector space. (c) Let W be the set of all matrices A in V such that A 21 = A 12 (the bar denotes complex conjugation). Prove that W is a subspace of V and find a basis for W. (3) Prove that the space of all m n matrices over the field F has dimension mn, by exhibiting a basis for this space.